2 * Intel SMP support routines.
4 * (c) 1995 Alan Cox, Building #3 <alan@redhat.com>
5 * (c) 1998-99, 2000 Ingo Molnar <mingo@redhat.com>
7 * This code is released under the GNU General Public License version 2 or
11 #include <linux/init.h>
14 #include <linux/irq.h>
15 #include <linux/delay.h>
16 #include <linux/spinlock.h>
17 #include <linux/smp_lock.h>
18 #include <linux/kernel_stat.h>
19 #include <linux/mc146818rtc.h>
20 #include <linux/cache.h>
21 #include <linux/interrupt.h>
24 #include <asm/tlbflush.h>
26 #include <mach_apic.h>
29 * Some notes on x86 processor bugs affecting SMP operation:
31 * Pentium, Pentium Pro, II, III (and all CPUs) have bugs.
32 * The Linux implications for SMP are handled as follows:
34 * Pentium III / [Xeon]
35 * None of the E1AP-E3AP errata are visible to the user.
42 * None of the A1AP-A3AP errata are visible to the user.
49 * None of 1AP-9AP errata are visible to the normal user,
50 * except occasional delivery of 'spurious interrupt' as trap #15.
51 * This is very rare and a non-problem.
53 * 1AP. Linux maps APIC as non-cacheable
54 * 2AP. worked around in hardware
55 * 3AP. fixed in C0 and above steppings microcode update.
56 * Linux does not use excessive STARTUP_IPIs.
57 * 4AP. worked around in hardware
58 * 5AP. symmetric IO mode (normal Linux operation) not affected.
59 * 'noapic' mode has vector 0xf filled out properly.
60 * 6AP. 'noapic' mode might be affected - fixed in later steppings
61 * 7AP. We do not assume writes to the LVT deassering IRQs
62 * 8AP. We do not enable low power mode (deep sleep) during MP bootup
63 * 9AP. We do not use mixed mode
66 * There is a marginal case where REP MOVS on 100MHz SMP
67 * machines with B stepping processors can fail. XXX should provide
68 * an L1cache=Writethrough or L1cache=off option.
70 * B stepping CPUs may hang. There are hardware work arounds
71 * for this. We warn about it in case your board doesn't have the work
72 * arounds. Basically thats so I can tell anyone with a B stepping
73 * CPU and SMP problems "tough".
75 * Specific items [From Pentium Processor Specification Update]
77 * 1AP. Linux doesn't use remote read
78 * 2AP. Linux doesn't trust APIC errors
79 * 3AP. We work around this
80 * 4AP. Linux never generated 3 interrupts of the same priority
81 * to cause a lost local interrupt.
82 * 5AP. Remote read is never used
83 * 6AP. not affected - worked around in hardware
84 * 7AP. not affected - worked around in hardware
85 * 8AP. worked around in hardware - we get explicit CS errors if not
86 * 9AP. only 'noapic' mode affected. Might generate spurious
87 * interrupts, we log only the first one and count the
89 * 10AP. not affected - worked around in hardware
90 * 11AP. Linux reads the APIC between writes to avoid this, as per
91 * the documentation. Make sure you preserve this as it affects
92 * the C stepping chips too.
93 * 12AP. not affected - worked around in hardware
94 * 13AP. not affected - worked around in hardware
95 * 14AP. we always deassert INIT during bootup
96 * 15AP. not affected - worked around in hardware
97 * 16AP. not affected - worked around in hardware
98 * 17AP. not affected - worked around in hardware
99 * 18AP. not affected - worked around in hardware
100 * 19AP. not affected - worked around in BIOS
102 * If this sounds worrying believe me these bugs are either ___RARE___,
103 * or are signal timing bugs worked around in hardware and there's
104 * about nothing of note with C stepping upwards.
107 DEFINE_PER_CPU(struct tlb_state, cpu_tlbstate) ____cacheline_aligned = { &init_mm, 0, };
110 * the following functions deal with sending IPIs between CPUs.
112 * We use 'broadcast', CPU->CPU IPIs and self-IPIs too.
115 static inline int __prepare_ICR (unsigned int shortcut, int vector)
117 return APIC_DM_FIXED | shortcut | vector | APIC_DEST_LOGICAL;
120 static inline int __prepare_ICR2 (unsigned int mask)
122 return SET_APIC_DEST_FIELD(mask);
125 void __send_IPI_shortcut(unsigned int shortcut, int vector)
128 * Subtle. In the case of the 'never do double writes' workaround
129 * we have to lock out interrupts to be safe. As we don't care
130 * of the value read we use an atomic rmw access to avoid costly
131 * cli/sti. Otherwise we use an even cheaper single atomic write
139 apic_wait_icr_idle();
142 * No need to touch the target chip field
144 cfg = __prepare_ICR(shortcut, vector);
146 if (vector == CRASH_DUMP_VECTOR)
147 cfg = (cfg&~APIC_VECTOR_MASK)|APIC_DM_NMI;
150 * Send the IPI. The write to APIC_ICR fires this off.
152 apic_write_around(APIC_ICR, cfg);
155 void fastcall send_IPI_self(int vector)
157 __send_IPI_shortcut(APIC_DEST_SELF, vector);
161 * This is only used on smaller machines.
163 void send_IPI_mask_bitmask(cpumask_t cpumask, int vector)
165 unsigned long mask = cpus_addr(cpumask)[0];
169 local_irq_save(flags);
174 apic_wait_icr_idle();
177 * prepare target chip field
179 cfg = __prepare_ICR2(mask);
180 apic_write_around(APIC_ICR2, cfg);
185 cfg = __prepare_ICR(0, vector);
188 * Send the IPI. The write to APIC_ICR fires this off.
190 apic_write_around(APIC_ICR, cfg);
192 local_irq_restore(flags);
195 void send_IPI_mask_sequence(cpumask_t mask, int vector)
197 unsigned long cfg, flags;
198 unsigned int query_cpu;
201 * Hack. The clustered APIC addressing mode doesn't allow us to send
202 * to an arbitrary mask, so I do a unicasts to each CPU instead. This
203 * should be modified to do 1 message per cluster ID - mbligh
206 local_irq_save(flags);
208 for (query_cpu = 0; query_cpu < NR_CPUS; ++query_cpu) {
209 if (cpu_isset(query_cpu, mask)) {
214 apic_wait_icr_idle();
217 * prepare target chip field
219 cfg = __prepare_ICR2(cpu_to_logical_apicid(query_cpu));
220 apic_write_around(APIC_ICR2, cfg);
225 cfg = __prepare_ICR(0, vector);
228 * Send the IPI. The write to APIC_ICR fires this off.
230 apic_write_around(APIC_ICR, cfg);
233 local_irq_restore(flags);
236 #include <mach_ipi.h> /* must come after the send_IPI functions above for inlining */
239 * Smarter SMP flushing macros.
240 * c/o Linus Torvalds.
242 * These mean you can really definitely utterly forget about
243 * writing to user space from interrupts. (Its not allowed anyway).
245 * Optimizations Manfred Spraul <manfred@colorfullife.com>
248 static cpumask_t flush_cpumask;
249 static struct mm_struct * flush_mm;
250 static unsigned long flush_va;
251 static DEFINE_SPINLOCK(tlbstate_lock);
252 #define FLUSH_ALL 0xffffffff
255 * We cannot call mmdrop() because we are in interrupt context,
256 * instead update mm->cpu_vm_mask.
258 * We need to reload %cr3 since the page tables may be going
259 * away from under us..
261 static inline void leave_mm (unsigned long cpu)
263 if (per_cpu(cpu_tlbstate, cpu).state == TLBSTATE_OK)
265 cpu_clear(cpu, per_cpu(cpu_tlbstate, cpu).active_mm->cpu_vm_mask);
266 load_cr3(swapper_pg_dir);
271 * The flush IPI assumes that a thread switch happens in this order:
272 * [cpu0: the cpu that switches]
273 * 1) switch_mm() either 1a) or 1b)
274 * 1a) thread switch to a different mm
275 * 1a1) cpu_clear(cpu, old_mm->cpu_vm_mask);
276 * Stop ipi delivery for the old mm. This is not synchronized with
277 * the other cpus, but smp_invalidate_interrupt ignore flush ipis
278 * for the wrong mm, and in the worst case we perform a superflous
280 * 1a2) set cpu_tlbstate to TLBSTATE_OK
281 * Now the smp_invalidate_interrupt won't call leave_mm if cpu0
282 * was in lazy tlb mode.
283 * 1a3) update cpu_tlbstate[].active_mm
284 * Now cpu0 accepts tlb flushes for the new mm.
285 * 1a4) cpu_set(cpu, new_mm->cpu_vm_mask);
286 * Now the other cpus will send tlb flush ipis.
288 * 1b) thread switch without mm change
289 * cpu_tlbstate[].active_mm is correct, cpu0 already handles
291 * 1b1) set cpu_tlbstate to TLBSTATE_OK
292 * 1b2) test_and_set the cpu bit in cpu_vm_mask.
293 * Atomically set the bit [other cpus will start sending flush ipis],
295 * 1b3) if the bit was 0: leave_mm was called, flush the tlb.
296 * 2) switch %%esp, ie current
298 * The interrupt must handle 2 special cases:
299 * - cr3 is changed before %%esp, ie. it cannot use current->{active_,}mm.
300 * - the cpu performs speculative tlb reads, i.e. even if the cpu only
301 * runs in kernel space, the cpu could load tlb entries for user space
304 * The good news is that cpu_tlbstate is local to each cpu, no
305 * write/read ordering problems.
311 * 1) Flush the tlb entries if the cpu uses the mm that's being flushed.
312 * 2) Leave the mm if we are in the lazy tlb mode.
315 fastcall void smp_invalidate_interrupt(struct pt_regs *regs)
320 if (current->active_mm)
321 load_user_cs_desc(cpu, current->active_mm);
323 if (!cpu_isset(cpu, flush_cpumask))
326 * This was a BUG() but until someone can quote me the
327 * line from the intel manual that guarantees an IPI to
328 * multiple CPUs is retried _only_ on the erroring CPUs
329 * its staying as a return
334 if (flush_mm == per_cpu(cpu_tlbstate, cpu).active_mm) {
335 if (per_cpu(cpu_tlbstate, cpu).state == TLBSTATE_OK) {
336 if (flush_va == FLUSH_ALL)
339 __flush_tlb_one(flush_va);
344 smp_mb__before_clear_bit();
345 cpu_clear(cpu, flush_cpumask);
346 smp_mb__after_clear_bit();
348 put_cpu_no_resched();
351 static void flush_tlb_others(cpumask_t cpumask, struct mm_struct *mm,
356 * A couple of (to be removed) sanity checks:
358 * - we do not send IPIs to not-yet booted CPUs.
359 * - current CPU must not be in mask
360 * - mask must exist :)
362 BUG_ON(cpus_empty(cpumask));
364 cpus_and(tmp, cpumask, cpu_online_map);
365 BUG_ON(!cpus_equal(cpumask, tmp));
366 BUG_ON(cpu_isset(smp_processor_id(), cpumask));
370 * i'm not happy about this global shared spinlock in the
371 * MM hot path, but we'll see how contended it is.
372 * Temporarily this turns IRQs off, so that lockups are
373 * detected by the NMI watchdog.
375 spin_lock(&tlbstate_lock);
379 #if NR_CPUS <= BITS_PER_LONG
380 atomic_set_mask(cpumask, &flush_cpumask);
384 unsigned long *flush_mask = (unsigned long *)&flush_cpumask;
385 unsigned long *cpu_mask = (unsigned long *)&cpumask;
386 for (k = 0; k < BITS_TO_LONGS(NR_CPUS); ++k)
387 atomic_set_mask(cpu_mask[k], &flush_mask[k]);
391 * We have to send the IPI only to
394 send_IPI_mask(cpumask, INVALIDATE_TLB_VECTOR);
396 while (!cpus_empty(flush_cpumask))
397 /* nothing. lockup detection does not belong here */
402 spin_unlock(&tlbstate_lock);
405 void flush_tlb_current_task(void)
407 struct mm_struct *mm = current->mm;
411 cpu_mask = mm->cpu_vm_mask;
412 cpu_clear(smp_processor_id(), cpu_mask);
415 if (!cpus_empty(cpu_mask))
416 flush_tlb_others(cpu_mask, mm, FLUSH_ALL);
420 void flush_tlb_mm (struct mm_struct * mm)
425 cpu_mask = mm->cpu_vm_mask;
426 cpu_clear(smp_processor_id(), cpu_mask);
428 if (current->active_mm == mm) {
432 leave_mm(smp_processor_id());
434 if (!cpus_empty(cpu_mask))
435 flush_tlb_others(cpu_mask, mm, FLUSH_ALL);
440 void flush_tlb_page(struct vm_area_struct * vma, unsigned long va)
442 struct mm_struct *mm = vma->vm_mm;
446 cpu_mask = mm->cpu_vm_mask;
447 cpu_clear(smp_processor_id(), cpu_mask);
449 if (current->active_mm == mm) {
453 leave_mm(smp_processor_id());
456 if (!cpus_empty(cpu_mask))
457 flush_tlb_others(cpu_mask, mm, va);
462 static void do_flush_tlb_all(void* info)
464 unsigned long cpu = smp_processor_id();
467 if (per_cpu(cpu_tlbstate, cpu).state == TLBSTATE_LAZY)
471 void flush_tlb_all(void)
473 on_each_cpu(do_flush_tlb_all, NULL, 1, 1);
477 * this function sends a 'reschedule' IPI to another CPU.
478 * it goes straight through and wastes no time serializing
479 * anything. Worst case is that we lose a reschedule ...
481 void smp_send_reschedule(int cpu)
483 send_IPI_mask(cpumask_of_cpu(cpu), RESCHEDULE_VECTOR);
486 void crash_dump_send_ipi(void)
488 send_IPI_allbutself(CRASH_DUMP_VECTOR);
492 * Structure and data for smp_call_function(). This is designed to minimise
493 * static memory requirements. It also looks cleaner.
495 static DEFINE_SPINLOCK(call_lock);
497 struct call_data_struct {
498 void (*func) (void *info);
505 static struct call_data_struct * call_data;
508 * this function sends a 'generic call function' IPI to all other CPUs
512 int smp_call_function (void (*func) (void *info), void *info, int nonatomic,
515 * [SUMMARY] Run a function on all other CPUs.
516 * <func> The function to run. This must be fast and non-blocking.
517 * <info> An arbitrary pointer to pass to the function.
518 * <nonatomic> currently unused.
519 * <wait> If 1, wait (atomically) until function has completed on other CPUs.
520 * If 0, wait for the IPI to be received by other CPUs, but do not wait
521 * for the completion of the function on each CPU.
522 * If -1, do not wait for other CPUs to receive IPI.
523 * [RETURNS] 0 on success, else a negative status code. Does not return until
524 * remote CPUs are nearly ready to execute <<func>> or are or have executed.
526 * You must not call this function with disabled interrupts or from a
527 * hardware interrupt handler or from a bottom half handler.
530 static struct call_data_struct dumpdata;
531 struct call_data_struct normaldata;
532 struct call_data_struct *data;
533 int cpus = num_online_cpus()-1;
538 /* Can deadlock when called with interrupts disabled */
539 /* Only if we are waiting for other CPU to ack */
540 WARN_ON(irqs_disabled() && wait >= 0);
542 spin_lock(&call_lock);
544 /* if another cpu beat us, they win! */
546 spin_unlock(&call_lock);
555 atomic_set(&data->started, 0);
556 data->wait = wait > 0 ? wait : 0;
558 atomic_set(&data->finished, 0);
563 /* Send a message to all other CPUs and wait for them to respond */
564 send_IPI_allbutself(CALL_FUNCTION_VECTOR);
566 /* Wait for response */
568 while (atomic_read(&data->started) != cpus)
572 while (atomic_read(&data->finished) != cpus)
574 spin_unlock(&call_lock);
579 static void stop_this_cpu (void * dummy)
584 cpu_clear(smp_processor_id(), cpu_online_map);
586 disable_local_APIC();
587 if (cpu_data[smp_processor_id()].hlt_works_ok)
588 for(;;) __asm__("hlt");
593 * this function calls the 'stop' function on all other CPUs in the system.
596 void smp_send_stop(void)
598 smp_call_function(stop_this_cpu, NULL, 1, 0);
601 disable_local_APIC();
606 * Reschedule call back. Nothing to do,
607 * all the work is done automatically when
608 * we return from the interrupt.
610 fastcall void smp_reschedule_interrupt(struct pt_regs *regs)
615 fastcall void smp_call_function_interrupt(struct pt_regs *regs)
617 void (*func) (void *info) = call_data->func;
618 void *info = call_data->info;
619 int wait = call_data->wait;
623 * Notify initiating CPU that I've grabbed the data and am
624 * about to execute the function
627 atomic_inc(&call_data->started);
629 * At this point the info structure may be out of scope unless wait==1
637 atomic_inc(&call_data->finished);